Hey, I’m Abhinav.
I work on AI for chip design. I’m interested in what happens when models meet real engineering: the tools, the tests, and the hardware underneath it all.
A few things I’ve worked on
01 / Selected work
At AMD’s Applied AI Labs, I work on coding-assistant evaluation, simulation-based feedback, and tools that help designers make sense of waveforms. The interesting question is how well an agent holds up on actual chip-design work.
View full poster ↗With Professor Parmesh Ramanathan, I studied how GPU-style soft errors affect semantic segmentation, including safety-relevant classes like cars and people.
The UW research symposium ↗VRSight explores AI-driven scene descriptions for VR accessibility. The work includes a full paper and a CHI demonstration.
Find the research on Scholar ↗I help run SiliconBadgers, where students get hands-on with digital design, verification, and the chip-development process. We’re working on an open-source LLM inference accelerator, with teams tackling software, architecture, RTL, verification, and physical design.
I help shape the technical direction and turn it into concrete work for each team, from profiling llama.cpp to defining interfaces and building verification infrastructure.
Learning the whole path
from models to hardware.
ML systems / Interactive explainer
GPU throughput, training time, cloud pricing, and a calculator to see how the assumptions change the answer.
Read the note & try the estimator04 / Learning resources
My practical guides to Synopsys on UW–Madison CAE. Real screenshots, terminal commands, GUI walkthroughs, and the code behind them.
Explore the resources ↗05 / A little context
My background is in computer engineering and computer science at UW–Madison. Before AMD, I worked in Professor Daifeng Wang’s lab on machine learning for genomics: large GPU clusters, large datasets, and efficient model adaptation.
That brought me to the questions I keep coming back to: how we train and serve models, how we know an agent is doing useful work, and what the hardware is doing underneath.
The résumé version ↗Things I keep coming back to
Agents become interesting when they can act, get useful feedback, and try again. I care about the evaluations and engineering tools that make that loop meaningful.
Training, inference, data movement. I’m drawn to the practical work of making model stacks run reliably and efficiently at scale.
The model is one layer. The compilers, memory, and compute units underneath it decide how performance lands in the real world.